Table 1.
Main results on mouse benchmarks
| Method | Pre-trained weights | Data ratio | mAP | RMSE | Dataset | Architecture |
|---|---|---|---|---|---|---|
| Zero-shot | SuperAnimal | – | 50.397 | 14.32 | DLC_Openfield | DLCRNet |
| Zero-shot | SuperAnimal | – | 95.219 | 4.881 | DLC_Openfield | HRNetw32 |
| Zero-shot | SuperAnimal | – | 96.348 | 4.572 | DLC_Openfield | AnimalTokenPose |
| Transfer learning | ImageNet | 0.01 | 62.226 | 18.136 | DLC_Openfield | DLCRNet |
| Transfer learning | ImageNet | 0.01 | 91.458 | 7.001 | DLC_Openfield | HRNetw32 |
| Transfer learning | ImageNet | 1.00 | 99.23 | 2.340 | DLC_Openfield | DLCRNet |
| Transfer learning | ImageNet | 1.00 | 100 | 1.131 | DLC_Openfield | HRNetw32 |
| Memory replay | SuperAnimal | 0.01 | 74.225 | 7.688 | DLC_Openfield | DLCRNet |
| Memory replay | SuperAnimal | 0.01 | 99.599 | 2.381 | DLC_Openfield | HRNetw32 |
| Memory replay | SuperAnimal | 1.00 | 97.946 | 3.071 | DLC_Openfield | DLCRNet |
| Memory replay | SuperAnimal | 1.00 | 99.868 | 1.210 | DLC_Openfield | HRNetw32 |
| Zero-shot | SuperAnimal | – | 76.139 | 9.013 | TriMouse | HRNetw32 |
| Zero-shot | SuperAnimal | – | 70.372 | 10.580 | TriMouse | AnimalTokenPose |
| Transfer learning | ImageNet | 0.01 | 26.116 | 31.562 | TriMouse | HRNetw32 |
| Transfer learning | ImageNet | 1.00 | 97.730 | 2.276 | TriMouse | HRNetw32 |
| Memory replay | SuperAnimal | 0.01 | 90.320 | 5.850 | TriMouse | HRNetw32 |
| Memory replay | SuperAnimal | 1.00 | 98.547 | 2.103 | TriMouse | HRNetw32 |
The mAP on multiple architectures, CNN (HRNet, DLCRNet), and Transformer based models (AnimalTokenPose model) on SuperAnimal-TopViewMouse. As a reminder, transfer learning means using a randomly initialized decoder that is also trained. Memory replay involves fine-tuning the encoder and decoder.